Ubuntu Intel Performance Still In Bad Shape

[phoronix]

Phoronix: Ubuntu Intel Performance Still In Bad Shape

We began talking about Intel graphics regressions in Ubuntu 9.04 back in January but for the most part that went under the radar at Canonical up until Ubuntu 9.04 was nearing release. At that point it was then explored whether greedy migration heuristics improved performance as the UXA acceleration architecture was still too problematic to enable by default. We had found that using some of the latest kernel code had improved the performance some, but still there were major regressions within Intel's new Linux driver stack.

http://www.phoronix.com/vr.php?view=13800

[MaestroMaus]

Why do you even look at karmic right now? It has been 13 days since the release of 9.04. You can not expect anything from it yet.

[VisezTrance]

I've noticed some more serious regressions on Fedora, affecting intel 915gm and 945gm chipsets. The release of fedora 11 will mark one year since these problems occurred.
It's only obvious Ubuntu will suffer from them as well.

[wswartzendruber]

I'm thinking Intel might be backing porting the DRI drivers over to Gallium3D.

[macemoneta]

Fedora recently built:

kernel-2.6.29.2-129.fc11.x86_64
mesa-dri-drivers-7.5-0.9.fc11.x86_64
mesa-libGL-7.5-0.9.fc11.x86_64
mesa-libGL-devel-7.5-0.9.fc11.x86_64
mesa-libGLU-7.5-0.9.fc11.x86_64
mesa-libGLU-devel-7.5-0.9.fc11.x86_64
xorg-x11-drv-intel-2.7.0-4.fc11.x86_64

With those levels, kernel mode setting and UXA are very stable (on my G45/X4500HD), and performance is good. Compiz still has some issues, but stability is improving quickly at this point.

[wswartzendruber]

Fedora recently built:

kernel-2.6.29.2-129.fc11.x86_64
mesa-dri-drivers-7.5-0.9.fc11.x86_64
mesa-libGL-7.5-0.9.fc11.x86_64
mesa-libGL-devel-7.5-0.9.fc11.x86_64
mesa-libGLU-7.5-0.9.fc11.x86_64
mesa-libGLU-devel-7.5-0.9.fc11.x86_64
xorg-x11-drv-intel-2.7.0-4.fc11.x86_64

With those levels, kernel mode setting and UXA are very stable (on my G45/X4500HD), and performance is good. Compiz still has some issues, but stability is improving quickly at this point.
Sweet! You have Gallium3D enabled?

[Sarvatt]

The opengl problem started happening with mesa master (7.6) around 5/5 and there were no problems running any of those on 5/4, you just happened to catch a transient bug when you decided to benchmark :( Such is the life of daily builds though :D Of course karmic's linux-libc-dev is broken right now and most packages wont build to update and see if it's fixed. If anything I would recommend benchmarking jaunty's UXA (instead of EXA that you used) vs karmic's UXA with xorg-edgers drivers though to see actual performance increase/decreases, perhaps even using mesa 7.5 in edger's jaunty section instead of 7.6 so all tests can be run. UXA in general has a big performance hit on all of my GMA950 hardware.. I don't game on intel GMA though so thats the least of my worries.. :) UXA only just stabilized enough to be used full time on my 945GME aspire one on may 2nd.

[macemoneta]

Sweet! You have Gallium3D enabled?

As far as I know Gallium3D is not yet enabled in Mesa 7.5.

[jbrown96]

I've come to expect shoddy reporting on this site. The articles are just an excuse to link to your own content. Most of the sentences don't even make sense because you're trying to link every story that you've ever written about a topic. There's no discussion of what's actually happening. There are a few "paragraphs" at the beginning of the article, which are mostly included for linking.
Then there's the ubiquitous graphs throughout that just take up space. Honestly, just say, "performance regressed 13% on test X from 9.04 to 9.10." I'm perfectly capable of reading. My favorite part is the explanation of some of the graphs. After a graph that reads 1329.31 for 9.04 and 906.15 for 9.10, we get some fabulous insight: "The lines composition performance had regressed." Or how about this: "The performance with the transformed blit bilinear operation was nearly the same." If you take out the test name (and associated grammatical necessities it introduces, which is six words), the explanation is a whopping six words, of which none are insightful.

Who is the target the audience? If its new users to the site (who presumably wouldn't know much about Linux graphics benchmarks) then why the dependence on graphs that don't make much sense? What's a JXRenderMark?? If the audience is experienced users who visit the site, why the constant re-hashing of everything you've ever done on the topic? People who read the site at all know about these regressions, and there's no need to link to 16 Phoronix articles (vs. two non-Phoronix articles). I understand a section at the end for further reading, but there's no reason it needs to take up article space. I guess it makes sense when you realize that the article is 623 words, of which 63 are hyperlinked text. I wish I could insert a graph... but I'll also tell you that it's just over 10% of the words, which is one in 10.

I think that this one really takes it to a new low; as another poster pointed out, it's been two weeks since 9.04 was released. Combine that with an inflammatory headline, and there is no value in this "article."

Phoronix has really gone overboard with its test suite. Who cares about this stuff? There's nothing here that matters to anyone. Even if performance has regressed, which is presumably important to (future) users, there's no discussion about why it has or what can be done to minimize it. Any system in a computer is extraordinarily complex, especially something like graphics that is integrated at so many parts; yet, there's no technical information. No code analysis. No quotes or views from developers. There really isn't anything but a couple pages of graphs with painfully obvious commentary.

I understand this is going to piss a lot of people off, but there's no denying its true.

This is over the line, but I simply can't resist. James Downey (from Billy Madison) said, "what you've just said is one of the most insanely idiotic things I have ever heard. At no point in your rambling, incoherent response were you even close to anything that could be considered a rational thought. Everyone in this room is now dumber for having listened to it. I award you no points, and may God have mercy on your soul."

I've been reading this site for probably about 1.5 years. This article is not unique; I usually bite my tongue and don't say anything; I don't like to mean to people and ridicule them, but at some point, it's just too much. This post is very negative, and I'm not going to apologize for that. Maybe it's a little too blunt, but we're all adults. That being said, there is some decent stuff on this site. I just wish that there wasn't such an effort to constantly publish articles, even when there's nothing to report.

As an ending note, I thought that it would be humorous to point out the length of this post: 670 words.

[Ant P.]

Well said.

[wswartzendruber]

Regardless, every kernel version since 2.6.28 has seen improvements in GEM's speed, and this will speed up UXA. As far as OpenGL goes, let's hold off until the end of the year when Gallium3D is more mainstream.

[DanL]

Yeah, Phoronix usually goes a bit overboard with pimping its test suite, but given its business model, I can usually ignore the shameless self-promotion. However, I do agree that this article is fairly useless.

[ernstp]

I've come to expect shoddy reporting on this site...

Well we come here anyway, don't we?! :-)

[ernstp]

I think KMS is the key here, Ubuntu needs to enable that. Let's see some new benchmarks then.

[chaos386]

I'm a bit disappointed with Phoronix here, too. I've experienced a near 6x improvement (http://www.phoronix.com/forums/showthread.php?p=72698#post72698) in 3D gaming performance on my G43, but this article only has one basic 3D test among a bunch of 2D ones. They also only tested one Intel IGP. Had they tested more chipsets and done more 3D tests, maybe there could have been some insight (i.e. if Intel is focusing more on 3D and newer chipsets at the expense of 2D performance on older chips). With such a narrow view, I'm not really surprised Canonical didn't pay much attention to Phoronix's warnings of performance regressions.

[MaestroMaus]

Well we come here anyway, don't we?! :-)
He wrote, that he did not dislike all articles, did he?

Same story for me, I skim the articles because sometimes there is something good to be found here.

[7oby]

I agree with most points jbrown96 mentioned. And I also agree with what chaos386 said: One IGP doesn't tell anything about the very different and widespread GM965/GM45 counterparts.

Furthermore I'd like to give you an example of how to collapse your benchmark diagrams in the future:
http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2795&p=4

How come that the Phoronix Testsuite now features 100+ Tests (http://www.phoronix.com/scan.php?page=news_item&px=NzI0NQ), but you happen to post the most irrelevant ones? How exactly is a KDE user or even Gnome Firefox user affected by GTK rendering performance? Please give me an example of an application that is bottlenecked by GtkRadioButton rendering performance?

I serioulsy encourage to review the published benchmarks in terms of practical relevance. If you lack discussions of driver internals, there's not much point in posting lowlevel benchmarks and I suggest skipping that. This extends to the meaningless test of RamSpeed in different kernel versions (http://www.phoronix.com/scan.php?page=article&item=linux_2629_benchmarks&num=7).

To say something positive: I enjoyed the 32 vs. 64 Bit Linux articles once.

[kxmas]

JBrown, what would like the article authors to write between the graphs? There's no great wisdom in those brief sentences, but there isn't any wild speculation about the cause of any discrepancies either. I'm thankful that my time isn't wasted by reading pages of bullcrap.

The selection of tests causes much more concern for me. Rarely is an article posted where I'm not left wondering how a few other things perform.

[Sarvatt]

I think KMS is the key here, Ubuntu needs to enable that. Let's see some new benchmarks then.

KMS is easily enabled in karmic now, there's absolutely no performance difference between UMS and KMS in UXA for me though.

[damentz]

jbrown96, it would be more suitable for someone like you to switch to decaf. In the conclusion of your argument, I gathered that you can count pretty high without the aid of your hands and feet. Complaining about his verbose word choice or his graph descriptions isn't helping.

Michael is making it painfully obvious that the performance regression on intel REALLY SUCKS. The more press this issue gets, possibly the more faster it will be resolved.

Dicks like you are regressive to this fast nature of spreading word, stop posting long meaningless articles that are longer than your victim's.

[Sarvatt]

Here's my results vs the ones in this news post using 2.7.1rc1 drivers because I think its fairer to compare exa vs exa and I dont see ubuntu moving to the UXA only 2.8 series drivers anytime soon with all of the problems it has.. Our hardware is basically the same, only I have 512mb ram more. Also I'm using mesa 7.5rc1 instead of 7.6 that we're tracking in xorg-edgers which has some major problems right now.

http://global.phoronix-test-suite.com/index.php?k=profile&u=robert-27839-18675-31010

as I mentioned in the "Testing latest -intel in Ubuntu" thread, I set up another PPA besides xorg-edgers with what I consider more stable packages that I used for that test because xorg-edgers is tracking the absolute latest bleeding edge stuff which of course is prone to breaking.. this is just temporary until things stabilize a little more though.

https://launchpad.net/~sarvatt/+archive/sarvatt-graphics


edit: here's an updated one with KMS/UXA also.

http://global.phoronix-test-suite.com/index.php?k=profile&u=robert-26126-19706-24854